A large-scale heterogeneous task scheduling system for industrial agents

Abstract

The present application belongs to the technical field of industrial scheduling, and provides a large-scale heterogeneous task scheduling system for industrial agents, comprising: a cooperative scheduling network construction module: obtaining a device list and a to-be-scheduled production task, constructing a task agent group and a resource agent group; a scheduling understanding acquisition module: when a task agent issues a scheduling request, a resource agent generates corresponding scheduling understanding information according to real-time states and local optimization targets, and forms a scheduling understanding list; a scheduling understanding analysis module: performing consistency checking on the scheduling understanding information in the scheduling understanding list. The differentiated cognition of different devices in the industrial field to the same task is converted into comparable data with a unified format, effectively optimizing the problem that the scheduling decisions of multiple agents conflict with each other and fall into negotiation deadlock due to different local perspectives, and providing a complete data basis for conflict arbitration.

Description

Technical Field

[0001] This invention belongs to the field of industrial scheduling technology, specifically a large-scale heterogeneous task scheduling system for industrial intelligent agents. Background Technology

[0002] With the in-depth development of artificial intelligence manufacturing, modern industrial production sites are showing large-scale heterogeneous characteristics: on the one hand, the production line is equipped with a variety of heterogeneous equipment, including industrial robots, CNC machine tools, automated guided vehicles, intelligent warehousing equipment, and testing instruments. These devices have fundamental differences in terms of communication protocols, data interfaces, computing power, physical motion characteristics, and processing accuracy. On the other hand, the production tasks themselves are characterized by multiple varieties, small batches, and strong temporal coupling, and dynamic disturbances such as emergency order insertion, sudden equipment failure, and material supply delays occur frequently.

[0003] However, in actual industrial production, because each agent only grasps its own local state and follows its own preset local optimization goals, when facing the same global production task, they often generate conflicting scheduling interpretations based on their own perspectives. This leads to serious disagreements at the decision-making level. Specifically, different equipment agents propose mutually exclusive suggestions on the allocation direction, timing arrangement, or execution parameters of the same sub-task, causing the system to be unable to output unified scheduling instructions, resulting in repeated negotiations or decision deadlock. The reason for this is that existing technology systems generally lack the ability to deeply mine and systematically analyze massive amounts of historical industrial operating data when facing such multi-agent scheduling conflicts. They have failed to fully utilize industrial data mining and industrial data statistical methods to effectively track and evaluate the historical accuracy of long-term scheduling suggestions from various resource agents. As a result, the adjudication layer cannot rely on big data analysis to distinguish the superiority or inferiority of different equipment's predictive capabilities. If a suggestion from one party is forcibly adopted, the rejected party will not receive compensation, and its local state will continue to deviate from expectations, which will exacerbate conflicts in subsequent scheduling cycles and even cause secondary problems such as equipment deterioration and material backlog. At the same time, there are uncertainties such as sensor noise and equipment performance drift in the industrial field, and deviations will inevitably occur between actual execution and scheduling expectations.

[0004] Therefore, the present invention provides a large-scale heterogeneous task scheduling system for industrial intelligent agents. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, at least one technical problem raised in the background art is solved.

[0006] The technical solution adopted by this invention to solve its technical problem is: a large-scale heterogeneous task scheduling system for industrial intelligent agents, comprising:

[0007] Collaborative scheduling network construction module: Obtains a list of devices and production tasks to be scheduled, and constructs task agent groups and resource agent groups;

[0008] Scheduling understanding acquisition module: When the task agent issues a scheduling request, the resource agent generates corresponding scheduling understanding information based on the real-time status and local optimization objectives, forming a scheduling understanding list;

[0009] The scheduling understanding and analysis module performs consistency checks on the scheduling understanding information in the scheduling understanding list. If there are conflicts, it analyzes the long-term accuracy of the scheduling suggestions based on the historical scheduling records of each resource agent, generates a time-dimensional credibility score, and evaluates the intensity of the spatial cascading impact based on the spatial topology location information in the device list, generating a spatial-dimensional perturbation score.

[0010] The adjudication and instruction generation module: Based on the time dimension credibility score and the spatial dimension disturbance score, it comprehensively sorts the scheduling understanding information, determines the priority target group, and generates a formal scheduling instruction set;

[0011] Iterative scheduling module: When executing the formal scheduling instruction set, it collects the real-time execution parameters of each target device in real time, calculates the degree of deviation between the parameters and the expected execution parameters, and if the deviation exceeds the limit, it iteratively corrects the scheduling instructions using the real-time execution parameters as input.

[0012] The beneficial effects of this invention are as follows:

[0013] This invention constructs a multi-agent collaborative scheduling network that includes task agent groups and resource agent groups. Each resource agent generates structured scheduling understanding information based on its real-time status and local optimization objectives. This transforms the differentiated perceptions of different devices on the same task in the industrial field into comparable data with a unified format. This effectively optimizes the problem of conflicting scheduling decisions and negotiation deadlock caused by different local perspectives of multiple agents, and provides a complete data foundation for conflict resolution.

[0014] This invention performs consistency checks on the scheduling understanding list through a scheduling understanding analysis module. When a conflict is detected, it initiates a spatiotemporal dual-dimensional analysis, generating a time-dimensional reliability score based on historical scheduling records and a spatial-dimensional disturbance score based on spatial topology location information. This allows for the evaluation of the historical prediction accuracy of each resource agent and the cascading impact of scheduling schemes on downstream devices. The adjudication and instruction generation module further employs the Pareto optimal ranking method to select the optimal set of schemes that are mutually exclusive in terms of both time accuracy and spatial impact. Simultaneously, it generates compensation instructions for rejected resource agents to balance the load of each device in the system, thus resolving the potential for the rejected party to exacerbate subsequent scheduling conflicts due to a lack of compensation over a long period.

[0015] Each module of this invention is deployed locally on the edge scheduling node. The generation of scheduling understanding information, conflict analysis, adjudication, and instruction issuance are all completed at the edge. Data synchronization with the cloud is only performed when needed, reducing the transmission latency and bandwidth pressure caused by uploading massive amounts of data from the industrial site to the cloud. At the same time, the iterative scheduling module collects the actual execution parameters of the equipment in real time and compares them with the expected parameters. When the deviation exceeds the limit, the rescheduling process is automatically triggered, forming an edge closed-loop iterative mechanism. This enables the scheduling system to adaptively correct the equipment execution deviation, reduce the amplification of a single deviation along the process link, and improve the real-time performance, reliability, and robustness of task scheduling in a large-scale heterogeneous industrial environment. Attached Figure Description

[0016] The invention will now be further described with reference to the accompanying drawings.

[0017] Figure 1 This is a system block diagram of an embodiment of the present invention;

[0018] Figure 2 This is a flowchart of the steps in an embodiment of the present invention;

[0019] Figure 3 This is an example diagram of the suggested production scheduling time distribution in the scheduling understanding and acquisition module of this embodiment of the invention;

[0020] Figure 4 This is an example diagram illustrating the deviation analysis of atomic subtask S-001 in the iterative scheduling module of this invention embodiment;

[0021] Figure 5 This is a flowchart of the judgment process of the scheduling understanding and analysis module in an embodiment of the present invention. Detailed Implementation

[0022] To make the technical means, creative features, objectives and effects of this invention easier to understand, the invention will be further described below in conjunction with specific embodiments.

[0023] Example

[0024] Please see Figure 1 - Figure 5 As shown in the embodiment of the present invention, a large-scale heterogeneous task scheduling system for industrial intelligent agents includes:

[0025] Collaborative scheduling network construction module: Obtain the equipment list of multiple heterogeneous devices participating in scheduling in the industrial site and the set of production tasks to be scheduled, decompose the production tasks into multiple atomic sub-tasks, and construct task intelligent agent groups and resource intelligent agent groups;

[0026] In this module, the collaborative scheduling network construction module is deployed in the edge scheduling node at the workshop level or production line level. It is configured to acquire the devices of multiple heterogeneous devices participating in the scheduling in the industrial site locally at the edge. All resource agents corresponding to all devices constitute a resource agent group.

[0027] Receive a set of production tasks to be scheduled from the Manufacturing Execution System (MES) or Enterprise Resource Planning System (ERP);

[0028] Each production task must contain at least the following structured fields: unique task identifier, product model or process route number, total task quantity, delivery time window, priority label: used to mark categories such as urgent orders, regular orders or trial production orders, and associated process document index;

[0029] For example, the MES system issues a production task, and the structured fields are shown in Table 1:

[0030] Fields value illustrate Task unique identifier T-20260713-008 Batch unique number Product Model / Process Route Number PistonRod-A (Piston Rod Type A) Corresponding process document number: PRC-A-032 Total number of tasks 200 pieces 200 piston rods need to be machined Delivery time window Earliest start time: 08:00 on the same day; latest completion time: 16:00 on the same day. 8-hour delivery window Priority tags Urgent orders The customer requested delivery before 16:00 on the same day. Related process document index PRC-A-032 Pointing to the process chain description in the digital twin library

[0031] Table 1

[0032] The production task is broken down into atomic subtasks, as follows:

[0033] Based on the associated process document index, the complete process specification document corresponding to the product model is retrieved from the digital twin library. The process specification document records all the processing steps that the product needs to go through from blank to finished product in the form of structured data, the process parameters of each step, the equipment selection requirements, the time quota, and the sequential constraints between steps. The process sequence corresponding to the product is extracted from the process document.

[0034] Extract the process sequence from the process specification document, and perform the following atomization process on each process:

[0035] Generate a globally unique atomic subtask identifier for the current process. The identifier is composed of the parent task identifier and the process number according to a fixed encoding rule.

[0036] Write the equipment type required for the current process into the equipment type requirement field of the atomic subtask;

[0037] Write the current process accuracy requirement into the accuracy requirement field of the atomic subtask;

[0038] Multiply the single-piece time or batch time of the current process by the total task quantity to calculate the standard total processing time of the process, and write it into the estimated time field of the atomic subtask. For processes that are timed per piece, the standard total processing time is equal to the single-piece time multiplied by the total task quantity. For processes that are processed in batches, the standard total processing time is directly adopted from the batch time specified in the process specification document.

[0039] Extract the list of tooling, cutting tools, and auxiliary materials required for the current process from the process specification document and write it into the resource requirement list field of the atomic subtask;

[0040] According to the process specification document, the atomic subtask identifiers corresponding to all the preceding processes of the current process are written into the pre-dependency set field of the atomic subtask. For the first process, the pre-dependency set is an empty set.

[0041] Based on the delivery time window of the parent task, combined with the standard total processing time of each process and the sequence constraints between processes, the start time window and the latest completion time window of each process are calculated using a backward calculation method and written into the time window field of the atomic subtask.

[0042] Specifically, the latest completion time of the last process is equal to the latest completion time of the parent task, and the start time is equal to the latest completion time minus the standard total processing time of the process. For any process that is not the last process, the latest completion time is equal to the start time of the next process minus the preset transfer buffer time between processes, and the start time is equal to its latest completion time minus the standard total processing time of the process.

[0043] Inherit the priority label of the parent task to all atomic subtasks to which the parent task belongs, so that each atomic subtask carries the same priority category as the parent task;

[0044] After performing the above atomization process on all processes in sequence, all atomic subtasks corresponding to the production task are generated;

[0045] All task agents corresponding to all atomic subtasks constitute a task agent group.

[0046] Scheduling understanding acquisition module: When the task agent issues a scheduling request, the resource agent generates corresponding scheduling understanding information based on its real-time state and local optimization goals, forming a scheduling understanding list;

[0047] In this module, the scheduling understanding and acquisition module is deployed in the edge scheduling node and is configured to acquire the scheduling request issued by the task agent locally at the edge and send the scheduling request message to all adjacent resource agents of the task agent through the bidirectional communication link established between the task agent and the resource agent.

[0048] After each resource agent receives a scheduling request message from each task agent through a communication link, it generates scheduling understanding information;

[0049] Based on its internal status, including current operating status, equipment type, capacity parameters, current load, and maintenance plan, the resource agent determines whether it can meet the equipment type requirements, accuracy requirements, and capacity requirements of the task within the requested time window. If any condition is not met, the resource agent generates only one rejection scheduling understanding message for this scheduling request, which includes the identifier of the rejected atomic subtask and the reason for rejection.

[0050] If all conditions are met, then adoptive scheduling understanding information is generated for the current scheduling request, including:

[0051] The resource agent reads the estimated processing time and available start time window from the scheduling request, and at the same time reads the current load timeline in its own internal state. The later of the cumulative end time of the current load and the start time of the requested available start time window is used as the suggested production start time.

[0052] Add the estimated processing time to the suggested production start time to get the suggested production completion time.

[0053] Compare the suggested production completion time with the latest completion time in the scheduling request;

[0054] If the suggested production completion time is earlier than or equal to the latest completion time, it will be marked as ready for on-time delivery;

[0055] If the suggested completion time is later than the latest completion time, the delay time is calculated, which is equal to the suggested completion time minus the latest completion time.

[0056] The estimated energy consumption value is obtained by multiplying the estimated processing time, the rated power of the equipment, and the power correction factor of the equipment at the corresponding load rate.

[0057] The power correction factor reflects the ratio between the actual operating power and the rated power of the equipment under non-full load conditions, and its value ranges from 0.3 to 1.0.

[0058] Read the average energy consumption value of the corresponding equipment under the historical same period operating conditions. The historical same period operating conditions refer to the historical processing records that are the same as or similar to the processing parameters (including processing materials, accuracy requirements, and estimated duration range) of the current scheduling request.

[0059] If the expected energy consumption is less than or equal to the average energy consumption, the energy efficiency score will be recorded as 1.

[0060] If the expected energy consumption is greater than the average energy consumption, the energy efficiency score will be recorded as the ratio of the average energy consumption to the expected energy consumption.

[0061] Read the adjacent device list field in its internal state to obtain the identifiers of all adjacent devices that have a direct upstream or downstream relationship with the device.

[0062] For each downstream adjacent device, the suggested production scheduling completion time is added to the standard transmission time from this device to the downstream adjacent device, which is read from the process correlation diagram, to obtain the expected material arrival time.

[0063] Among them, the process association map is a directed graph data structure pre-stored locally on the edge scheduling node. It takes all the equipment in the workshop as nodes and the material flow direction between the equipment as directed edges, recording the direct upstream and downstream connection relationships between each piece of equipment.

[0064] Read the current load timeline of downstream adjacent equipment to determine whether the expected material arrival time falls within the busy load period of downstream adjacent equipment.

[0065] The busy load interval is set based on the time interval occupied by all tasks that have been scheduled but not yet completed in the current load timeline of downstream adjacent equipment, with the planned start time and planned completion time of the task as the boundary.

[0066] If it does not fall within the peak hours, it is marked as having no risk of congestion;

[0067] If it falls into the busy period, it is marked as having a congestion risk, and the expected congestion duration is calculated. The expected congestion duration is equal to the remaining completion time of the tasks being performed by the downstream adjacent equipment at the expected material arrival time.

[0068] The analysis results of all downstream adjacent devices are summarized into a cascaded effect declaration array;

[0069] Each element in the cascading impact statement array contains the downstream adjacent device identifier, the expected material arrival time, the congestion risk indicator, and the expected congestion duration (if any).

[0070] All the above information is summarized and encapsulated into complete scheduling understanding information. The understanding information includes at least the following fields: a unique identifier of the device corresponding to the resource agent that issued the understanding information; an atomic subtask identifier of the scheduling request corresponding to the understanding information; a suggested production start time; a suggested production completion time; an on-time delivery indicator or delay duration; an energy efficiency score; an array of cascading effect declarations; and a timestamp at the time of generation.

[0071] All scheduling understanding information is constructed into a scheduling understanding list.

[0072] In the scheduling understanding and acquisition module, the resource agent uses the real-time synchronized current load timeline as the data basis to calculate the suggested production scheduling time. It uses the real-time load timeline of downstream adjacent equipment as the judgment basis to generate a cascading impact statement. At the same time, driven by the local optimization objective, it adopts the scheduling strategy of prioritizing the earliest idle time to calculate the suggested production scheduling scheme. By comparing the expected energy consumption with the historical average energy consumption, it generates an energy efficiency score. This makes the output scheduling understanding information reflect the actual operating conditions of the equipment and also embed the decision preferences of each resource agent to maximize its own utilization and optimize energy consumption, thus providing a decision data source for the subsequent adjudication module.

[0073] Its function is to transform the differentiated decision preferences and operational condition perception results of various resource agents for the same scheduling request into scheduling understanding information with a unified format, thereby providing a complete, structured and horizontally comparable decision data source for the subsequent spatiotemporal dual-dimensional analysis module to perform conflict detection and comprehensive adjudication.

[0074] Scheduling Understanding and Analysis Module: Performs consistency checks on all scheduling understanding information in the scheduling understanding list. If conflicts exist, it performs spatiotemporal analysis.

[0075] The spatiotemporal analysis includes: analyzing the long-term accuracy of scheduling suggestions based on the historical scheduling records of each resource agent, and generating a time-dimensional credibility score for each scheduling understanding information; and evaluating the intensity of spatial cascading effects caused by each scheduling understanding information based on the spatial topology location information in the device list, and generating a spatial-dimensional perturbation score.

[0076] In this module, the scheduling understanding and analysis module is deployed in the edge scheduling node and is configured to perform consistency verification on all scheduling understanding information in the scheduling understanding list passed by the scheduling understanding acquisition module locally at the edge. When there is a conflict, spatiotemporal dual-dimensional analysis is started locally at the edge to generate a time dimension credibility score and a spatial dimension perturbation score for each scheduling understanding information.

[0077] Specifically, a consistency check is performed on all scheduling understanding information, including:

[0078] Check whether the unique identifiers of devices in each scheduling understanding message are different. If two scheduling understanding messages come from the same resource agent, they are judged as data anomalies and removed from the scheduling understanding list.

[0079] Check for conflicts in the scheduling information;

[0080] Among them, the conflict situations include at least the following: there is no common idle time period among the devices corresponding to each scheduling understanding information to accommodate the estimated processing time of the atomic subtask, that is, all candidate devices are unable to complete the task within the requested time window; the suggested production scheduling schemes of multiple scheduling understanding information cannot avoid time overlap with the already scheduled tasks on the time axis of their respective devices.

[0081] If there is no conflict, the one with the highest energy efficiency score will be selected and sent to the priority decision module.

[0082] If a conflict exists, a spatiotemporal analysis will be performed;

[0083] It should be noted that if the total number of scheduling understanding information corresponding to the same atomic subtask is only one, there is no need to perform conflict analysis. The scheduling understanding information is directly output as the only candidate solution for the atomic subtask to the priority decision module.

[0084] Specifically, the spatiotemporal analysis process includes:

[0085] For each piece of scheduling understanding information, locate the historical scheduling record storage area of ​​the corresponding resource agent from the local cache of the edge scheduling node, and read all the historical scheduling records of the resource agent in the past preset historical time period.

[0086] Among them, the historical scheduling records should at least include the suggested production completion time from the scheduling understanding information made by the resource agent in each instance, as well as the actual execution completion time of the corresponding task.

[0087] In the historical scheduling records, extract the number of records in which the absolute value of the deviation between the suggested production completion time and the actual execution completion time does not exceed the preset deviation tolerance threshold, and use them as accurate records;

[0088] The historical accuracy rate is calculated by comparing the accurate records with the total number of historical records. The historical accuracy rate is then used as a credibility score for the time dimension.

[0089] The deviation tolerance threshold is set based on the statistical distribution characteristics of the deviation between the suggested production completion time and the actual execution completion time in the historical scheduling records of each device. The percentile or standard deviation multiple of the deviation value sequence is used as the threshold setting benchmark. At the same time, the administrator can manually adjust the deviation according to the leniency of the workshop production rhythm and the strict requirements of order delivery.

[0090] It should be noted that when the resource agent does not have historical scheduling records, the time dimension credibility score is set to the default median value, which is 0.5.

[0091] For each piece of scheduling understanding information, read the cascading impact declaration array in the scheduling understanding information. The cascading impact declaration array contains the resource agent's expected material arrival time, congestion risk indicator, and expected congestion duration for each adjacent downstream device.

[0092] The number of devices at risk of congestion is counted as the disturbance quantity. The disturbance quantity is then calculated as a ratio to the total number of devices to obtain the disturbance quantity percentage.

[0093] The larger the number of disturbances, the more downstream devices the current scheduling information may affect simultaneously.

[0094] Extract the expected congestion durations for all devices at risk of congestion and sum them up. Then, combine the sum with the total number of downstream adjacent devices to obtain the average congestion duration.

[0095] Extract the expected congestion duration corresponding to all devices with congestion risk, take the maximum value as the maximum disturbance duration, and calculate the ratio between the average congestion duration and the maximum disturbance duration to obtain the disturbance duration ratio;

[0096] Among them, the longer the average congestion duration, the more severe the interference of the current scheduling understanding information on the downstream production rhythm.

[0097] The spatial dimension perturbation score is obtained by arithmetically averaging the percentage of perturbation quantity with the percentage of perturbation duration.

[0098] In the scheduling understanding and analysis module, the original scheduling understanding information is received, and consistency verification is performed locally on multiple candidate solutions for the same atomic subtask at the edge. The scheduling conflicts caused by different resource agents' working condition perception and decision-making preferences are identified. Through time dimension analysis (generating a credibility score based on the historical prediction accuracy of each resource agent) and spatial dimension analysis (generating a disturbance score based on the number of congested devices and the average congestion duration in the cascading impact statement), the differentiated device understanding that is difficult to compare directly is transformed into data with a unified ranking basis. This enables the priority decision module to rank multiple candidate solutions according to the credibility score and the disturbance score, and select the scheduling solution with the most accurate prediction and the least impact on downstream processes.

[0099] The adjudication and instruction generation module comprehensively sorts all scheduling understanding information in the understanding list based on the time dimension credibility score and the spatial dimension disturbance score, determines the priority target group, and generates a formal scheduling instruction set based on the scheduling understanding information corresponding to the priority target group.

[0100] Meanwhile, for resource agents that are not included in the priority target group and whose scheduling understanding information is rejected, compensation instructions are generated. These compensation instructions are used to adjust load balancing or maintain priority in subsequent scheduling cycles.

[0101] In this module, based on the time-dimensional credibility score and the spatial-dimensional perturbation score, the Pareto dominance method is used to perform optimal ranking. The process is as follows:

[0102] Each piece of scheduling understanding information is used as a decision-making scheme, and each scheme contains two decision dimensions: time dimension credibility score and spatial dimension perturbation score;

[0103] Among them, a higher credibility score in the time dimension indicates that the historical predictions of the resource agent are more accurate; a higher disturbance score in the spatial dimension indicates that the scheduling scheme has a greater negative impact on downstream equipment.

[0104] Define any two scheduling interpretation messages corresponding to the same atomic subtask as Scheme A and Scheme B respectively, and determine the dominance relationship according to the following rules:

[0105] Option A dominates Option B if and only if the time dimension reliability score of Option A is greater than or equal to the time dimension reliability score of Option B, and the spatial dimension perturbation score of Option A is less than or equal to the spatial dimension perturbation score of Option B, and at least one of the above two inequalities is a strict inequality (i.e., when Option A is not inferior to Option B in the time dimension, not inferior to Option B in the spatial dimension, and strictly superior to Option B in at least one dimension), then Option A is said to Pareto dominate Option B.

[0106] Traverse all scheduling understanding information. For each scheduling understanding information, check whether there are other scheduling understanding information that Pareto dominate the current solution. If a dominance relationship exists, the dominated solution is removed from the candidate set. After all traversals and removals, all remaining scheduling understanding information is the Pareto optimal solution set.

[0107] In the Pareto optimal solution set, any two scheduling solutions are independent of each other: if a solution is better in the time dimension, it must be worse in the space dimension; conversely, if a solution is better in the space dimension, it must be worse in the time dimension.

[0108] When the Pareto optimal solution set contains only one scheduling understanding piece of information, the corresponding resource agent is directly identified as the priority target of the current atomic subtask.

[0109] When the Pareto optimal solution set contains multiple scheduling understanding information, a pre-defined tie-breaking rule is used to determine the single priority objective from among them;

[0110] The tie-breaking rule is as follows: in the Pareto optimal solution set, the solution with the lowest spatial dimension perturbation score is selected as the priority target; if the spatial dimension perturbation scores are still the same, the solution with the highest time dimension credibility score is selected as the priority target.

[0111] Traverse all atomic subtasks and aggregate the priority target resource agents corresponding to each atomic subtask to form a priority target group;

[0112] It should be noted that the priority target group refers to the set of priority target resource agents determined by Pareto optimal sorting for each of the atomic subtasks in the current scheduling batch.

[0113] Based on the scheduling understanding information corresponding to each resource agent in the priority target group, a formal scheduling instruction set is generated;

[0114] A formal scheduling instruction set shall include at least the following: a unique identifier for each atomic subtask; a unique identifier for the assigned priority target equipment; a suggested start time; a suggested completion time; and specific processing parameters and process requirements.

[0115] As a preferred embodiment, for resource agents that are not included in the priority target group and whose scheduling understanding information is rejected, i.e., resource agents whose scheduling understanding information is dominated and eliminated in the Pareto optimal ranking process or who are not selected in the tie-breaking rule, a corresponding compensation instruction is generated.

[0116] Specifically, iterate through all scheduling understanding information for all atomic subtasks, and for the resource agent corresponding to each scheduling understanding information that was not selected as a priority target, perform the following compensation decision:

[0117] Read the suggested production start time and estimated processing time from the scheduling understanding information rejected by the resource agent. If the equipment is idle during the corresponding time period because the suggested production plan of the resource agent is not adopted, the idle time period is recorded as the capacity loss of the resource agent.

[0118] Based on the capacity loss, capacity compensation is given to the resource agent in subsequent scheduling cycles;

[0119] The specific method of capacity compensation is as follows: In the equipment load timeline of the resource agent, idle time periods are marked as compensation quotas. In future scheduling cycles, when a new task arrives and the equipment type requirement of the task matches that of the resource agent, the resource agent has priority to be scheduled under the same conditions until the compensation quota is completely consumed.

[0120] In the adjudication and instruction generation module, the Pareto optimal sorting method is used to select a set of Pareto optimal solutions that are mutually exclusive in terms of both time prediction accuracy and spatial impact. By breaking the tie rule, the optimal solution for each atomic subtask is determined from these solutions. These solutions are then aggregated to form a priority target group and generate a formal set of dispatchable instructions. This enables each physical device to execute production tasks according to the adjudication results. At the same time, capacity compensation instructions are generated for resource agents that are not adopted. Idle periods caused by non-scheduling are recorded as priority quotas for subsequent scheduling, reducing the load imbalance of devices in the system caused by long-term neglect. This achieves optimized management of scheduling decisions.

[0121] Iterative scheduling module: When executing the formal scheduling instruction set, it collects the real-time execution parameters of each target device in real time, calculates the degree of deviation between the parameters and the expected execution parameters, and if the deviation exceeds the limit, it iteratively corrects the scheduling instructions using the real-time execution parameters as input.

[0122] In this module, during the execution of the formal scheduling instruction set, the actual execution parameters of each target device are continuously collected;

[0123] The actual execution parameters include at least: the current operating status of the equipment, the execution time of the current machining task, the estimated remaining completion time of the current machining task, and the real-time operating data of key components of the equipment (including but not limited to the current spindle speed, current cutting load, and current cumulative tool usage time).

[0124] Align the real-time execution parameters with the corresponding expected execution parameters at the execution time;

[0125] Calculate the absolute value of the difference between the real-time execution parameters and the corresponding expected execution parameters, as the degree of deviation;

[0126] The degree of deviation is compared with a preset deviation threshold.

[0127] Among them, the deviation threshold is the maximum allowable deviation range corresponding to each type of execution parameter, which is preset by the system administrator according to the production cycle requirements of the workshop, the stringency of order delivery, and the fluctuation range of the equipment's own performance.

[0128] When the deviation is less than or equal to the deviation threshold, the current task execution progress is determined to be within an acceptable range, and monitoring continues.

[0129] When the deviation exceeds the deviation threshold, the current task execution progress is determined to be out of acceptable range, triggering a rescheduling process.

[0130] The rescheduling process is as follows: using the actual execution parameters of all target devices collected at the current moment as the new device status input, using all unfinished atomic subtasks as the new set of tasks to be scheduled, calling the collaborative scheduling network construction module to the adjudication and instruction generation module again, executing the complete scheduling process, and generating the corrected scheduling instruction set;

[0131] The revised scheduling instruction set is sent to the equipment that has not yet completed the corresponding task for execution. At the same time, confirmation and maintenance instructions are sent to the equipment that has completed the task as originally planned to continue to execute the subsequent scheduled tasks. During the rescheduling process, the completed atomic subtasks do not participate in the new scheduling process, and only the atomic subtasks that have not yet been completed are rescheduled.

[0132] The above-mentioned collection, calculation, judgment and rescheduling process is continuously executed. At the preset sampling period, the actual execution parameters are continuously collected, the degree of deviation is calculated, whether the limit is exceeded and rescheduling is triggered until all atomic subtasks in the current scheduling batch have been completed.

[0133] As a specific implementation method of this embodiment: take the estimated remaining completion time of the current processing task as an example;

[0134] During the execution of the formal scheduling instruction set, the target device is currently executing atomic subtask S-001. The expected execution parameter of the atomic subtask in the scheduling instruction set is 10:00 AM on the day of the suggested production completion time. The actual execution parameters of the device are continuously collected at a preset sampling period (for example, the sampling period is one minute).

[0135] At a sampling moment (for example, the current time is 9:30 a.m.), the estimated remaining completion time is read from the actual execution parameters collected as 40 minutes. The current time (9:30 a.m.) is added to the estimated remaining completion time (40 minutes) to calculate the actual predicted completion time as 10:10 a.m.

[0136] The difference between the actual predicted completion time (10:10 AM) and the suggested production completion time of the atomic sub-task (10:00 AM) is calculated to obtain a time deviation value of ten minutes (10:10 AM minus 10:00 AM equals ten minutes). The absolute value of the time deviation value is used as the degree of deviation of the atomic sub-task at the current moment, that is, the degree of deviation is ten minutes.

[0137] The deviation level (ten minutes) is compared with a preset deviation level threshold. For example, the deviation level threshold is preset to five minutes. If ten minutes is greater than five minutes, it is determined that the current task execution progress exceeds the acceptable range and a rescheduling process is triggered.

[0138] The rescheduling process is as follows: using the actual execution parameters of all target devices collected at the current time (9:30 a.m. on the same day) as the new device status input, and using all the atomic subtasks that have not yet been completed (including the currently executing S-001 and all subsequent atomic subtasks that have not yet started) as the new set of tasks to be scheduled, the collaborative scheduling network construction module is called again to the adjudication and instruction generation module to execute the complete scheduling process and generate the corrected scheduling instruction set;

[0139] The revised scheduling instruction set is issued to the equipment that has not yet completed the corresponding task. For the target equipment, the revised suggested production completion time of atomic subtask S-001 in the revised scheduling instruction set is updated to 10:08 on the same day (the achievable completion time is recalculated based on the current actual progress). At the same time, a confirmation and maintenance instruction is sent to the equipment that has completed the task as originally planned, so that the subsequent scheduled tasks can continue to be executed. During the rescheduling process, the completed atomic subtasks do not participate in the new scheduling process, and only the atomic subtasks that have not yet been completed are rescheduled.

[0140] It should be noted that if the deviation calculated at a certain sampling time is less than or equal to the deviation threshold (for example, the estimated remaining completion time is thirty minutes, the actual predicted completion time is 10:00 on the same day, the time deviation value is zero minutes, and the deviation level is zero minutes), then the current task execution progress is determined to be within an acceptable range, rescheduling is not triggered, and monitoring continues.

[0141] In the iterative scheduling module, during the execution of the formal scheduling instruction set, the actual execution parameters of each target device are continuously collected and compared with the expected execution parameters in real time. Once the deviation exceeds the limit, the rescheduling process is triggered. The corrected scheduling instruction set is regenerated with the current actual working conditions as input, thus forming an edge closed-loop iterative mechanism of execution, monitoring, correction, and re-execution. This ensures that the scheduling system can adaptively correct the execution deviation of the equipment, reduce the chain reaction of delays in subsequent processes caused by the amplification of a single deviation along the process link, and prevent completed sub-tasks from participating in rescheduling to control the calculation until all sub-tasks in the current batch are completed. This realizes online dynamic correction of scheduling instructions and improves the robustness of the system.

[0142] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A large-scale heterogeneous task scheduling system for industrial intelligent agents, characterized in that: include: Collaborative scheduling network construction module: Obtains a list of devices and production tasks to be scheduled, and constructs task agent groups and resource agent groups; Scheduling understanding acquisition module: When the task agent issues a scheduling request, the resource agent generates corresponding scheduling understanding information based on the real-time status and local optimization objectives, forming a scheduling understanding list; The scheduling understanding and analysis module performs consistency checks on the scheduling understanding information in the scheduling understanding list. If there are conflicts, it analyzes the long-term accuracy of the scheduling suggestions based on the historical scheduling records of each resource agent, generates a time-dimensional credibility score, and evaluates the intensity of the spatial cascading impact based on the spatial topology location information in the device list, generating a spatial-dimensional perturbation score. The adjudication and instruction generation module: Based on the time dimension credibility score and the spatial dimension disturbance score, it comprehensively sorts the scheduling understanding information, determines the priority target group, and generates a formal scheduling instruction set; Iterative scheduling module: When executing the formal scheduling instruction set, it collects the real-time execution parameters of each target device in real time, calculates the degree of deviation between the parameters and the expected execution parameters, and if the deviation exceeds the limit, it iteratively corrects the scheduling instructions using the real-time execution parameters as input.

2. The large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The construction process of the task agent group and the resource agent group is as follows: All resource agents corresponding to all devices in the device list constitute a resource agent group. Based on the associated process document index, the corresponding complete process specification document is used to extract the process sequence from the process specification document, and each production task to be scheduled is broken down into multiple atomic sub-tasks according to the process sequence. All task agents corresponding to all atomic subtasks constitute a task agent group.

3. The large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The scheduling understanding information should include at least: the suggested production start time; the suggested production completion time; the on-time delivery indicator or the delay duration; the energy efficiency score; and the cascading effect declaration array. Read the estimated processing time and available start time window from the scheduling request, read the current load timeline from its own internal status, and use the latest of the current load's cumulative end time and the requested start time window start time as the suggested production start time. Add the estimated processing time to the suggested production start time to get the suggested production completion time. If the suggested production completion time is earlier than or equal to the latest completion time, it will be marked as ready for on-time delivery; If the suggested completion time is later than the latest completion time, the suggested completion time minus the latest completion time is calculated as the delay duration.

4. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 3, characterized in that: The process of obtaining the energy efficiency score is as follows: The estimated energy consumption value is obtained by multiplying the estimated processing time, the rated power of the equipment, and the power correction factor of the equipment at the corresponding load rate. Read the average energy consumption of the corresponding device under historical operating conditions; If the expected energy consumption is less than or equal to the average energy consumption, the energy efficiency score will be recorded as 1. If the expected energy consumption is greater than the average energy consumption, the energy efficiency score is recorded as the ratio of the average energy consumption to the expected energy consumption.

5. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 3, characterized in that: The process of obtaining the declared array for cascading effects is as follows: Obtain all adjacent devices that have a direct upstream or downstream relationship with the device; For each downstream adjacent device, the suggested production scheduling completion time will be added to the standard transmission time from this device to the downstream adjacent device to obtain the expected material arrival time. Read the current load timeline of downstream adjacent equipment; if the expected material arrival time does not fall within the busy period, mark it as having no congestion risk; if the expected material arrival time falls within the busy period, mark it as having congestion risk, and calculate the remaining completion time of the tasks being executed by downstream adjacent equipment at the expected material arrival time, as the expected congestion duration. The analysis results of all downstream adjacent devices are summarized into a cascaded effect declaration array.

6. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The process of performing consistency checks is as follows: Check whether the unique device identifiers in each scheduling understanding message are different. If two scheduling understanding messages come from the same resource agent, remove them from the scheduling understanding list. Check for conflicts in the scheduling information; If there is no conflict, the highest energy efficiency score will be used as the selection criterion; if there is a conflict, a spatiotemporal analysis will be performed.

7. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The process of obtaining the time dimension credibility score is as follows: Read all historical scheduling records of the resource agent within the preset historical time period; In the historical scheduling records, extract the number of records in which the absolute value of the deviation between the suggested production completion time and the actual execution completion time does not exceed the preset deviation tolerance threshold, and use them as accurate records; The historical accuracy rate is calculated by comparing the accurate records with the total number of historical records. This historical accuracy rate is then used as a reliability score for the time dimension.

8. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The process of obtaining the spatial dimension perturbation score is as follows: Read the cascading impact declaration array from the scheduling understanding information. The cascading impact declaration array contains the resource agent's expected material arrival time, congestion risk indicator, and expected congestion duration for each adjacent downstream device. The number of devices at risk of congestion is counted as the disturbance quantity. The disturbance quantity is then calculated as a ratio to the total number of devices to obtain the disturbance quantity percentage. Extract the expected congestion durations for all devices at risk of congestion and sum them up. Then, combine the sum with the total number of downstream adjacent devices to obtain the average congestion duration. Extract the expected congestion duration corresponding to all devices with congestion risk, take the maximum value as the maximum disturbance duration, and calculate the ratio between the average congestion duration and the maximum disturbance duration to obtain the disturbance duration ratio; The spatial dimension perturbation score is obtained by arithmetically averaging the percentage of perturbation quantity with the percentage of perturbation duration.

9. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The process of identifying the priority target group is as follows: Each piece of scheduling understanding information is used as a decision scheme, which includes two dimensions: time dimension credibility score and spatial dimension perturbation score. The Pareto dominance rule is used to eliminate the dominated schemes, and the remaining schemes constitute the Pareto optimal solution set. When there is only one solution in the Pareto optimal solution set, it is directly determined as the priority target. When there are multiple solutions, the one with the lowest spatial dimension perturbation score is selected. If the perturbation scores are the same, the one with the highest time dimension credibility score is selected. Traverse all atomic subtasks, summarize the priority target resource agents corresponding to each subtask, and form a priority target group.

10. A large-scale heterogeneous task scheduling system for industrial intelligent agents according to claim 1, characterized in that: The iterative scheduling module includes: The actual execution parameters of each target device are collected at a preset sampling period, and the degree of deviation between the actual parameters and the expected execution parameters is calculated. If the degree of deviation is less than or equal to a preset threshold, monitoring continues. If the value exceeds the preset threshold, the current actual execution parameters are used as input to re-invoke the collaborative scheduling network construction module to the adjudication and instruction generation module, generate a revised scheduling instruction set, and send it to the devices that have not yet completed their tasks for execution. Completed atomic subtasks do not participate in rescheduling, and the loop continues until all atomic subtasks are completed.

Images